# 📰 Governance Got Here First — Microsoft Agent 365 GA, Pentagon Reopens Talks, Korea's AI Center Goes Legal — 5/4 yyaia Brief *The signal is no longer model size. It is who builds the operating system that manages AI agents — and Microsoft just shipped one.* Issued · SF:// · 2026-05-04 (KST) --- ## 🧭 Editor's View — How to read today's headlines If we connect today's headlines into one line, **AI is now crossing from "tool" to "agent-as-OS."** The [general availability of Microsoft Agent 365](https://www.microsoft.com/en-us/security/blog/2026/05/01/microsoft-agent-365-now-generally-available-expands-capabilities-and-integrations/) is not just one more option for IT admins. The vast majority of humanity is still uncomfortable with computers — the share of people who actually use a computer in their daily life remains a minority. The moment an agent OS becomes standardized, that changes. A casual sentence — "shall we take a trip somewhere?" — finishes flights and hotel bookings on its own. **A world in which we use computers without looking at them.** That is the destination agentic voice has been pointing toward, and its first commercial infrastructure shipped on May 1. The question is the gap. The technology flow always reaches the center first. In the 1990s PC adoption wave, we had **facilitators** to teach how to use the machine, and that role narrowed the gap. Will agent OS just diffuse like smartphones did, or will it need another generation of facilitators? **What we do every week from the periphery — the technological shadow zones, the places where the flow of capital, infrastructure, and tooling arrives last — is exactly that facilitator's work.** Yeongyang is one anchor point we use to measure the gap; it stands for many similar places. Today's other line is the counterweight. [MIT's Recursive Language Models paper](https://arxiv.org/abs/2512.24601) shows small models can approach the quality of frontier ones. A window of opportunity has briefly opened for the underdogs. Its meaning needs to be measured precisely before the resource game closes it → **a dedicated deep review will follow in the next Academic Research section.** --- ## ⭐ Today's Big Story — Microsoft Agent 365 hits General Availability On May 1, Microsoft moved [Agent 365 out of preview into general availability](https://www.microsoft.com/en-us/security/blog/2026/05/01/microsoft-agent-365-now-generally-available-expands-capabilities-and-integrations/). Pricing is **$15 per user per month**, bundled inside the new **Microsoft 365 E7** SKU. The platform now syncs with **AWS Bedrock and Google Cloud agent registries**. IT teams can inventory and govern agents *across other clouds* from a single console — the first time that has been possible at this scale ([TechCrunch coverage](https://techcrunch.com/2026/05/01/pentagon-inks-deals-with-nvidia-microsoft-and-aws-to-deploy-ai-on-classified-networks/), among others). **Why this matters** — The defining question of 2026 is no longer "whose model is smarter." It is **who owns the operating system that runs the agents**. Microsoft has just standardized that control plane (the layer that supervises many agents from one place) in an Intune-style admin experience and shipped it commercially first. At the same release, support for **local AI agents via OpenClaw** went live. The same governance now reaches from cloud to desktop. > ⏱ **Time-to-arrival — center to periphery** > - Global big-tech IT (SF / Seoul / London): immediate (May 1) > - Mid-sized firms in Tier-2 cities: 6–12 months > - County-level government IT in non-metropolitan regions (Yeongyang, Cheongsong, Taebaek and the like): 2–3 years (E7 SKU adoption gap) > - Meaning: The first infrastructure of an "use-the-computer-without-looking-at-it" era has landed. From the periphery, our task is to walk over and touch the shape it will take when it gets here. --- ## 🚀 Models & Releases ### Google ships Gemini 3.1 Ultra and Flash-Lite together **Ultra** — 2-million-token context, native multimodal across text, image, audio and video. No transcription middleman; the model reads the video itself. **Flash-Lite** — $0.25 per million input tokens, 2.5× faster response, +45% output speed. The efficiency tier just dropped another floor. > The reading — two simultaneous moves. ① The **context-length race reignites at 2M.** ② The **low-end has a new $0.25 input price line.** ### DeepSeek V4 — price cut and long context This week's V4 lineup pushes the V3.2 baseline of **$0.14 / $0.28 per million tokens (input/output)** further down with a long-context option. GPT-4-class quality is now 7–10× cheaper than a year ago. --- ## 🔬 Research ### Recursive Language Models — MIT calls it "the paradigm of 2026" By [Alex L. Zhang · Tim Kraska · Omar Khattab](https://arxiv.org/abs/2512.24601). The core idea is simple — treat the long prompt as an *external environment*, then let the LM **recursively call itself**. The mechanism: context lives as a variable inside a Python REPL; the model dispatches a sub-LM (a smaller instance of itself) to specific segments only when needed. Crucially, **no summarization, no information loss**. Performance — RLMs handle inputs up to **two orders of magnitude beyond a model's context window**. RLM-Qwen3-8B approaches vanilla GPT-5 quality on three long-context benchmarks. Prime Intellect labels it **"the paradigm of 2026."** > **Why this matters** — A philosophical shift: it's not context **length** that matters, it's **management**. Small models just got a serious new path to imitating large ones. > > 📚 **Deep review forthcoming** — The SimThink engine — owned by **Myeongseong-Simjae Inc.**, an education company independent of SF:// — sits on top of recursive reasoning. Two opposing readings are at stake: small-model imitation as a window for underdogs, and the same paradigm running on top-tier compute as the next round of the resource game. We will treat both formally in the upcoming Academic Research section. ### Insilico Medicine ISM001-055 — first AI-discovered target with AI-designed drug to clear Phase IIa Idiopathic pulmonary fibrosis. [Phase IIa positive results](https://insilico.com/news/tnik-ipf-phase2a). The full study has been [published in Nature Medicine](https://www.nature.com/articles/s41591-025-03743-2). The **first time** an AI-discovered target plus an AI-designed molecule has shown human efficacy. > The reading — medical AI moves from "diagnostic assistant" to **drug discovery itself**. Q1 2026 digital-health VC funding was **$4B** (+$1B YoY); 12 mega-deals captured 59%. --- ## 🛠 Tools in the Field ### Microsoft Agent 365 — Bedrock + GCP agents on one console Same announcement as the Big Story, but the field-use angle: agents built in AWS Bedrock or Google Cloud Vertex are now auto-discovered and inventoried by Agent 365. A standard IT tool for **multi-cloud agent operations** has arrived. ### Microsoft DAX Copilot — adopted by 150+ U.S. health systems Clinical documentation automation has steepened its adoption curve. AI tooling has crossed from "feature" to "standard infrastructure" inside hospitals. --- ## 🏢 Industry & Business ### Pentagon × 7 vendors + Anthropic talks reopen Officially [signed May 1](https://www.war.gov/News/Releases/Release/Article/4475177/classified-networks-ai-agreements/): **SpaceX, OpenAI, Google, Microsoft, Nvidia, AWS, Reflection** — classified-network AI tooling. Oracle was added [hours later](https://winbuzzer.com/2026/05/03/pentagon-classified-ai-agreements-nvidia-microsoft-aws-google-openai-spacex-oracle-reflection-xcxwbn/). Anthropic was excluded after refusing to drop safety guardrails for fully autonomous weapons and mass surveillance use. In the same week: [White House Chief of Staff Susie Wiles and Treasury Secretary Bessent met Dario Amodei at the White House on April 17](https://www.cnbc.com/2026/04/17/anthropic-dario-amodei-trump-mythos.html). [Axios reports](https://www.axios.com/2026/04/17/anthropic-trump-administration-mythos) the meeting was "productive." Trump told CNBC a deal was "possible" and called the company "very smart." The **Mythos issue is treated separately**. Anthropic [remains on the Pentagon supply-chain risk list](https://www.axios.com/2026/04/29/trump-anthropic-pentagon-ai-executive-order-gov), but the channel is open again. > **Why this matters** — A frontier lab just showed the market that **walking away from defense revenue on safety grounds** is now a viable negotiating position. That has not happened before. ### Samsung Q1 chip operating profit ₩53.7T (×48 YoY) · SK hynix locks in ~70% of HBM4 [Samsung Electronics' semiconductor division reported a 48-fold Q1 OP surge to ₩53.7 trillion (~USD 36B)](https://www.tweaktown.com/news/111364/samsung-discloses-its-q1-2026-earnings-with-a-48x-growth-in-semiconductor-operating-profit/index.html). SK hynix posted a 5× quarterly profit jump and nearly tripled revenue. UBS estimates SK hynix has secured **~70% of initial HBM4 volume** for Nvidia's Vera Rubin (Nvidia's next-generation AI training/inference platform unveiled at GTC 2026). > The reading — Vera Rubin demands HBM4 at 10Gb/s+. **Memory is now the bottleneck of the next round.** Both Korean firms sit inside that bottleneck. --- ## ⚖️ Policy & Regulation ### California EO N-5-26 — Newsom directs AI safety draft (March 30) [On March 30, Governor Newsom signed Executive Order N-5-26](https://www.gov.ca.gov/2026/03/30/as-trump-rolls-back-protections-governor-newsom-signs-first-of-its-kind-executive-order-to-strengthen-ai-protections-and-responsible-use/). State agencies must produce AI safety recommendations. The package now covers AI safety and whistleblower protection, training-data transparency and watermarking, HR/ADS (automated decision systems) anti-discrimination, and a ban on "common pricing algorithms." California's stack thickens further. ### Colorado AI Law slips to June 30 Original Feb 1, 2026 → **June 30, 2026**. EU AI Act high-risk obligations are also under discussion to slip to 2027–2028. **Regulatory time keeps lagging market time.** --- ## 🇰🇷 Korea Focus ### National AI Computing Center — **SPC incorporated April 1**, groundbreaking July, 50K GPUs by 2030 A [consortium led by Samsung SDS, Samsung Electronics, Samsung C&T, Naver Cloud, Kakao, KT, plus one more](https://www.sedaily.com/NewsView/2K782VM4PB) — **legal entity established April 1**, July groundbreaking, 2029 opening, **50,000 GPUs by 2030**. Total project budget over **₩2 trillion (~USD 1.5B)**. > **Why this matters** — Korea's flagship AI infrastructure cleared the legal-entity hurdle in May. Fast for a public-sector project. But "by 2030" is still 4 years away from county-level peripheries — Yeongyang, Cheongsong, Jeongseon, Taebaek and similar places. ### Microsoft report — **Korea is #1 in 2025 AI adoption speed** Microsoft attributes it to a combination of government policy acceleration and the country's ChatGPT user base. > The arrival-time gap — the same gap we measure each week from the periphery. **#1 globally**, and yet still distant from the shadow zones. Yeongyang is one anchor point we use to mark that distance. ### Korea AI Association — **AI EXPO KOREA, COEX, May 6–8** The 9th International AI Expo in Seoul. The domestic industry's annual gathering point. ### Google I/O 2026 — confirmed for **May 19–20** Roadmap reveal expected; possible follow-ups to Gemini 3.1 Ultra and Flash-Lite. --- ## 🏛 Programs Applicable to Peripheral Regions (Yeongyang & Gyeongbuk as examples) **No new direct calls this week** for Yeongyang or Gyeongbuk. Three nationwide programs remain open and applicable to non-metropolitan SMEs and regional research institutes — including, but not limited to, places like Yeongyang and Cheongsong. 1. **Region-Led AI Transformation Program 2026** — open through Dec 18, 2026. Province-level proposals; Gyeongsangbuk-do may consortium. 2. **AX-Sprint** — 246 AI applied products, **₩754B (~USD 555M)**. Five domains: manufacturing, agri/livestock/fisheries, transport, health/welfare/environment, public-life security/defense. The **agri-livestock track is the most plausible entry point for county-level peripheries**. 3. **AI Compute Resource Strengthening** — **₩2.08T (~USD 1.5B)**, GPU resource allocation. Separate sub-tracks for regional research institutes and SMEs. > Periphery read — the **AX-Sprint agri/livestock track** is the closest entry. A Myeongseong-Simjae AI Academy × local farms consortium submitting a single applied product is a realistic simulation to walk through. The same template transfers cleanly to other county-level peripheries. --- ## ⚡ Quick Hits - **BMW i Ventures launches new $300M fund** — industrial AI weighting expanded (May 3). - **TurboQuant (Google, ICLR 2026)** — KV-cache compression via PolarQuant rotation + quantized Johnson-Lindenstrauss. Another inference-cost reduction path. - **KOCCA gaming program — up to ₩300M per AI-game project** — 2026 content track. Plausible extension lane for youth AI content cohorts in peripheral regions. - **MS Agent 365 + local AI agents (OpenClaw)** — cloud and desktop on the same governance plane. A solo operator gets the same interface as an enterprise. - **Microsoft DAX Copilot in 150+ U.S. health systems** — clinical documentation automation now standard. --- *— Chi-hoon Shin, Ph.D. · Editor* --- Licensed under CC BY-NC 4.0.